Non-negative Structured Pyramidal Neural Network for Pattern Recognition

Deep learning is a machine learning paradigm that has been widely exploited in the last years due to its high performance in many different problems, most of them related to computer vision. The Structured Pyramidal Neural Network (SPNN) is an artificial neural network which implements some of the deep learning concepts, such as multiple processing layers and receptive fields, with no need to pre-define the features of the problem as inputs. This network architecture has presented equivalent or even better results than other deep network approaches applied to solve specific tasks, but with a much lower computational cost. However, one of the SPNN limitations is the difficulty in contributing to human interpretations, since it has opaque learning, like most of the neural networks. Thus, we propose a non-negative model of the SPNN, to obtain better interpretability of the network learning. We restrict the values of the weights and biases of the network to be non-negative. The proposed model is evaluated in a gender recognition problem using the Face Recognition Technology (FERET) database. The results show that SPNN including non-negative constraint returns comparable recognition rates, but providing gains in the interpretability and stability of the model.